research · carlrannaberg — команда для Claude Code: что делает и как установить
Команда research разбивает вопрос на параллельные задачи субагентов, читает их отчёты из /tmp и сводит в один файл со ссылками на источники.
MIT ★ 763 обновлён 2026-03-31
.claude/commands/research.mdЧто делает research
Команда сначала относит вопрос к одному из трёх типов: широкий обзор, глубокая тема или простой факт. От этого зависит, сколько задач уйдёт в работу, от 1 до 10. Все они запускаются через Task в одном сообщении, поэтому субагенты идут одновременно, а не по очереди.
Промпт каждой задачи начинается с ключевого слова. Quick check: просит быструю проверку, Investigate: среднюю глубину, Deep dive: долгий разбор. Слово задаёт, сколько поисков сделает субагент. Полный отчёт он пишет в /tmp/research_[timestamp]_[topic].md, а назад возвращает путь, выжимку на пару предложений и число найденных источников.
Дальше основная сессия читает эти файлы через Read, убирает пересечения и складывает всё в /tmp/research_final_[timestamp].md: краткая выжимка, ключевые находки, разбор по темам, список источников. Инструменты ограничены Task, Read, Write, Edit, Grep и Glob, модель — sonnet. Задачи уходят с subagent_type="research-expert", так что субагент с таким именем должен быть определён в проекте.
Какие задачи решает
- сравнить несколько облачных провайдеров по независимым осям
- проверить дату релиза по первичным источникам
- разобрать одну техническую тему с нескольких сторон
- свести отчёты параллельных агентов в один файл
- убрать повторяющиеся ссылки из общего списка источников
Примеры запросов агенту
- /research Сравни три облачных провайдера по независимым осям
- /research Разбери, как в трансформерах устроен механизм внимания
- /research Проверь по первичным источникам, когда вышла Claude
Как установить research
Вручную: положить файлы
.claude/commands/research.mdВызов — /research. Для всех проектов — ~/.claude/commands/
Вручную: положить файлы
~/.codex/prompts/research.mdВызов — /prompts:research. В Codex этот формат помечен устаревшим, замена — скиллы
Вручную: положить файлы
.github/prompts/research.prompt.mdВручную: положить файлы
.gemini/commands/research.tomlGemini хранит команды в TOML: текст запроса — в поле prompt
Вопросы о research
Как установить research в Claude Code?
Положите файл команды в .claude/commands/research.md. Дальше вызывайте /research с вопросом, который нужно разобрать.
Что делает research?
Классифицирует вопрос, раскидывает его по параллельным задачам research-expert, а затем сводит их отчёты из /tmp в один файл с общим списком источников.
Что нужно, чтобы команда заработала?
В проекте должен быть определён субагент research-expert: задачи уходят с subagent_type="research-expert". Ещё команда работает на модели sonnet с набором инструментов из allowed-tools.
Исходный текст
research.md · MIT
--- description: Deep research with parallel subagents and automatic citations argument-hint: "<question to investigate>" allowed-tools: Task, Read, Write, Edit, Grep, Glob category: workflow model: sonnet --- # 🔬 Research Command Conduct deep, parallel research on any topic using multiple specialized subagents. ## Research Query $ARGUMENTS ## Research Process ### Phase 1: Query Classification (CRITICAL FIRST STEP) **PRIMARY DECISION: Classify the query type to determine research strategy** #### Query Types: 1. **BREADTH-FIRST QUERIES** (Wide exploration) - Characteristics: Multiple independent aspects, survey questions, comparisons - Examples: "Compare all major cloud providers", "List board members of S&P 500 tech companies" - Strategy: 5-10 parallel subagents, each exploring different aspects - Each subagent gets narrow, specific tasks 2. **DEPTH-FIRST QUERIES** (Deep investigation) - Characteristics: Single topic requiring thorough understanding, technical deep-dives - Examples: "How does transformer architecture work?", "Explain quantum entanglement" - Strategy: 2-4 subagents with overlapping but complementary angles - Each subagent explores the same topic from different perspectives 3. **SIMPLE FACTUAL QUERIES** (Quick lookup) - Characteristics: Single fact, recent event, specific data point - Examples: "When was GPT-4 released?", "Current CEO of Microsoft" - Strategy: 1-2 subagents for verification - Focus on authoritative sources #### After Classification, Determine: - **Resource Allocation**: Based on query type (1-10 subagents) - **Search Domains**: Academic, technical, news, or general web - **Depth vs Coverage**: How deep vs how wide to search ### Phase 2: Parallel Research Execution Based on the query classification, spawn appropriate research subagents IN A SINGLE MESSAGE for true parallelization. **CRITICAL: Parallel Execution Pattern** Use multiple Task tool invocations in ONE message, ALL with subagent_type="research-expert". **MANDATORY: Start Each Task Prompt with Mode Indicator** You MUST begin each task prompt with one of these trigger phrases to control subagent behavior: - **Quick Verification (3-5 searches)**: Start with "Quick check:", "Verify:", or "Confirm:" - **Focused Investigation (5-10 searches)**: Start with "Investigate:", "Explore:", or "Find details about:" - **Deep Research (10-15 searches)**: Start with "Deep dive:", "Comprehensive:", "Thorough research:", or "Exhaustive:" Example Task invocations: ``` Task(description="Academic research", prompt="Deep dive: Find all academic papers on transformer architectures from 2017-2024", subagent_type="research-expert") Task(description="Quick fact check", prompt="Quick check: Verify the release date of GPT-4", subagent_type="research-expert") Task(description="Company research", prompt="Investigate: OpenAI's current product offerings and pricing", subagent_type="research-expert") ``` This ensures all subagents work simultaneously AND understand the expected search depth through these trigger words. **Filesystem Artifact Pattern**: Each subagent saves full report to `/tmp/research_[timestamp]_[topic].md` and returns only: - File path to the full report - Brief 2-3 sentence summary - Key topics covered - Number of sources found ### Phase 3: Synthesis from Filesystem Artifacts **CRITICAL: Subagents Return File References, Not Full Reports** Each subagent will: 1. Write their full report to `/tmp/research_*.md` 2. Return only a summary with the file path Synthesis Process: 1. **Collect File References**: Gather all `/tmp/research_*.md` paths from subagent responses 2. **Read Reports**: Use Read tool to access each research artifact 3. **Merge Findings**: - Identify common themes across reports - Deduplicate overlapping information - Preserve unique insights from each report 4. **Consolidate Sources**: - Merge all cited sources - Remove duplicate URLs - Organize by relevance and credibility 5. **Write Final Report**: Save synthesized report to `/tmp/research_final_[timestamp].md` ### Phase 4: Final Report Structure The synthesized report (written to file) must include: # Research Report: [Query Topic] ## Executive Summary [3-5 paragraph overview synthesizing all findings] ## Key Findings 1. **[Major Finding 1]** - Synthesized from multiple subagent reports 2. **[Major Finding 2]** - Cross-referenced and verified 3. **[Major Finding 3]** - With supporting evidence from multiple sources ## Detailed Analysis ### [Theme 1 - Merged from Multiple Reports] [Comprehensive synthesis integrating all relevant subagent findings] ### [Theme 2 - Merged from Multiple Reports] [Comprehensive synthesis integrating all relevant subagent findings] ## Sources & References [Consolidated list of all sources from all subagents, organized by type] ## Research Methodology - Query Classification: [Breadth/Depth/Simple] - Subagents Deployed: [Number and focus areas] - Total Sources Analyzed: [Combined count] - Research Artifacts: [List of all /tmp/research_*.md files] ## Research Principles ### Quality Heuristics - Start with broad searches, then narrow based on findings - Prefer authoritative sources (academic papers, official docs, primary sources) - Cross-reference claims across multiple sources - Identify gaps and contradictions in available information ### Effort Scaling by Query Type - **Simple Factual**: 1-2 subagents, 3-5 searches each (verification focus) - **Depth-First**: 2-4 subagents, 10-15 searches each (deep understanding) - **Breadth-First**: 5-10 subagents, 5-10 searches each (wide coverage) - **Maximum Complexity**: 10 subagents (Claude Code limit) ### Parallelization Strategy - Spawn all initial subagents simultaneously for speed - Each subagent performs multiple parallel searches - 90% time reduction compared to sequential searching - Independent exploration prevents bias and groupthink ## Execution **Step 1: CLASSIFY THE QUERY** (Breadth-first, Depth-first, or Simple factual) **Step 2: LAUNCH APPROPRIATE SUBAGENT CONFIGURATION** ### Example Execution Patterns: **BREADTH-FIRST Example:** "Compare AI capabilities of Google, OpenAI, and Anthropic" - Classification: Breadth-first (multiple independent comparisons) - Launch 6 subagents in ONE message with focused investigation mode: - Task 1: "Investigate: Google's current AI products, models, and capabilities" - Task 2: "Investigate: OpenAI's current AI products, models, and capabilities" - Task 3: "Investigate: Anthropic's current AI products, models, and capabilities" - Task 4: "Explore: Performance benchmarks comparing models from all three companies" - Task 5: "Investigate: Business models, pricing, and market positioning for each" - Task 6: "Quick check: Latest announcements and news from each company (2024)" **DEPTH-FIRST Example:** "How do transformer models achieve attention?" - Classification: Depth-first (single topic, deep understanding) - Launch 3 subagents in ONE message with deep research mode: - Task 1: "Deep dive: Mathematical foundations and formulas behind attention mechanisms" - Task 2: "Comprehensive: Visual diagrams and step-by-step walkthrough of self-attention" - Task 3: "Thorough research: Seminal papers including 'Attention is All You Need' and subsequent improvements" **SIMPLE FACTUAL Example:** "When was Claude 3 released?" - Classification: Simple factual query - Launch 1 subagent with verification mode: - Task 1: "Quick check: Verify the official release date of Claude 3 from Anthropic" Each subagent works independently, writes findings to `/tmp/research_*.md`, and returns a lightweight summary. ### Step 3: SYNTHESIZE AND DELIVER After all subagents complete: 1. Read all research artifact files from `/tmp/research_*.md` 2. Synthesize findings into comprehensive report 3. Write final report to `/tmp/research_final_[timestamp].md` 4. Provide user with: - Executive summary (displayed directly) - Path to full report file - Key insights and recommendations **Benefits of Filesystem Artifacts**: - 90% reduction in token usage (passing paths vs full reports) - No information loss during synthesis - Preserves formatting and structure - Enables selective reading of sections - Allows user to access individual subagent reports if needed Now executing query classification and multi-agent research...
© carlrannaberg · MIT · carlrannaberg/claudekit
Похожие слэш-команды
Все слэш-команды по теме «Настройка агента: память, контекст, субагенты» →Сначала — research от других авторов.